AI 911 calls got believable enough this year to empty a California high school overnight, and the people who answer them cannot reliably hear the difference between a cloned voice and a terrified one. In August, a caller told Wheatland Union High School in Yuba County that a bomb was inside. Officers went in after dark, classes were canceled, and the building turned out to hold nothing but desks. The threat was ordinary. The voice was not.

Hoax calls to emergency lines are as old as the telephone. What changed in 2026 is the audio. A dispatcher used to be able to hear the seams in a fake: flat cadence, missing breath, a script read too evenly. That tell is gone, and the people who spent careers trusting it are the ones now being aimed at. The association that represents them says the technology is not merely making their calls harder. It is being pointed at them.

Mel Maier — The Voice Sounds Like a Witness Now, and That Is the Whole Problem

Mel Maier runs APCO International, the association that speaks for public safety communications professionals. His account of the change is blunt: the voices on these hoaxes used to be recognizable within seconds, and now they are not. The new ones hesitate. They breathe in the wrong places on purpose, drop in filler words, and pause as though the caller is deciding what to admit. Maier’s point is that they land, tonally, exactly where a nervous human would land — which means the one channel a call taker could trust has quietly closed.

That matters because swatting is not a prank with a victimless punchline. The call is designed to pull a large armed response toward a person the caller wants punished: a public official, an ex-partner, a stranger with the wrong handle on a gaming server. Every unit that rolls is a unit not rolling somewhere else. People have been killed in these raids, and the fatal hoax cases have produced federal prosecutions running to dozens of counts.

A liar used to have a tell. Now the tell is a toggle in a settings menu.

AI 911 Calls — $20 Buys the Voice, and 14% of Dispatchers Have Already Seen It

The scale is not speculative. APCO and the Public Safety Threat Alliance surveyed the field and published the findings in June: 766 public safety communications professionals responded, and roughly 14 percent said they had already been targeted with AI tools. Maier expects that share to climb as the tools improve, which is the least comforting forecast in public safety right now.

Brandon May, an assistant professor of forensic psychology at the Florida Institute of Technology, laid out the price of entry. Cloning his own voice on ElevenLabs cost him under twenty dollars, and the knobs included intonation, pitch, and the length of pauses. Twenty dollars does not buy a convincing impersonation of a specific person to a forensic examiner. It buys something more useful to a hoaxer: a voice that sounds like a neighbor in trouble. May’s own summary of the problem is that nobody can glance at an image or listen to a clip anymore and simply announce that it is synthetic — and the practical version of that sentence is that a call taker now has to make a life-or-death judgment with the audio evidence removed.

Fourteen percent is not a rounding error. It is a job hazard with a monthly subscription.

April Heinze — Swatters Stopped Dialing 911, Which Cut the Only Wire Worth Pulling

April Heinze, vice president and chief of 911 operations at NENA, the 9-1-1 Association, points at a detail most people never think about: the hoaxers have largely stopped calling 911 at all. Dialing 911 hands the center location data, a callback number, and the name on the account. Location on a 911 call is not spoofable. So the calls come in to a police department’s non-emergency line or a courthouse switchboard instead, and get transferred into the dispatch center from there — arriving stripped of the data that used to do the investigative work for free.

The result is a call that cannot be traced back to its origin, placed by a voice that cannot be outed by its sound, reporting an emergency that cannot be safely ignored. Three layers of verification, all removed by one routing choice. Heinze’s answer to the tracing question is that there is no tracing to do.

They did not beat the lock. They found the door nobody had bothered to shut.

The Dispatchers — They Will Send the Cars Anyway, and That Is the Point

Both associations are explicit about the operating rule inside a communications center: a report of a life-threatening situation is treated as real. Even when the indicators point to a hoax, the call taker sends the response, because the cost of being wrong in the other direction is a body. Heinze’s framing is that public safety will always answer hard, and the public should understand that the response itself is the thing being stolen.

Which puts the dispatcher in an impossible seat. The job already runs on a shrinking pool of people — the New Orleans center handles better than a thousand calls a day with roughly a third of its intake positions unfilled — and now every call carries an extra question that did not exist two years ago. To help, APCO is building interrogation criteria: question sets granular enough that a model cannot answer them from public records, a social feed, or a guess. Both Heinze and Maier declined to describe the specifics in public. Teaching the public the test is teaching the software the answer.

Publish the answer key and you have published the workaround.

AI 911 Dispatch — A Listener Is Already on the Line, Grading the Humans

The other half of this story is that the same technology is now sitting inside the dispatch center, on the city’s payroll. Long Beach confirmed it is rolling out CommsCoach AI, a GovWorx product that will listen to and score all 600,000 calls the city’s police and fire dispatch receives each year. It replaces the supervisor spot-checks that were supposed to cover about 12,000 calls and, in a short-staffed shop, often did not. The contract runs roughly $68,000 a year through 2028. Reginald Harrison, the city’s director of disaster preparedness, says there is no plan to replace dispatchers, and the dispatchers’ union was consulted before the council vote.

He is probably telling the truth, and it does not fully settle the question. The system flags calls where the tone, the phrasing, or a missing follow-up question looks wrong, and a supervisor reviews the flags. A dispatcher’s worst thirty seconds — a caller screaming, a language barrier, a decision made in four seconds — is now a transcript with a score attached. The same cheap voice cloning that defeats the human ear will be measured against software bought to grade the human response.

The city bought a listener. Nobody sold it a lie detector.

The Caller — The Person Who Really Is in Danger Pays the Bill

Here is the cost that does not show up in a survey. In a center running fourteen to sixteen people around the clock, every hoax consumes minutes that are not refundable. The response that answers a fake bomb threat is drawn out of a real shift. And the person on the other end who genuinely needs help — the one who called because someone was in the house, because the chest pain would not stop — now reaches a call taker who has been trained, by necessity, to interrogate before believing.

That is the trade the AI-mekna made on the way into the emergency line, and it is not reversible by policy. The dispatchers are not going to stop answering. They are going to keep sending the cars, keep asking the sharper questions, and keep absorbing the abuse of a system where the voice in their ear can no longer be trusted. The containment systems worked right up until they didn’t — and this time the bottle was opened by a phone call.

Sources: NPR, APCO International, Long Beach Post, Newsweek.